DocumentLevel MultiAspect Sentiment Classification for Online Reviews of

Document-Level Multi-Aspect Sentiment Classification for Online Reviews of Medical Experts ADVISOR: JIA-LING KOH SOURCE: CIKM 19 SPEAKER: LI-WEI Li. U DATA: 2020/05/11 1

OUTLINE l. Introduction l. Method l. Experiment l. Conclusion 2

INTRODUCTION Online doctor review system: Provide a platform for doctors and patients to operate and select Patient comments on the doctor https: //www. ratemds. com/ 3

INTRODUCTION Goal: üAutomatically score through patient comments on doctors üFor multiple aspects respectively Document-Level Multi-Aspect Sentiment Classification(DLMASC) 4

INTRODUCTION Document-Level Multi-Aspect Sentiment Classification(DLMASC) Output: Rating for each aspect Input: A review (Patient comments on the docto) 5

OUTLINE l. Introduction l. Method l. Experiment l. Conclusion 6

FRAMEWORK Multiple tasks, shared parameters predict rating features of doctors reviews aspect-keywords How to get? keywords for aspect k 7

METHOD Discover Aspect-Keywords • Latent Dirichlet Allocation(LDA) – Unsupervised Find the most relevant top-k keywords 8

METHOD Discover Aspect-Keywords • In our dataset… 9

METHOD for aspect k • Review Encoder Input: review X(x 1, x 2, …, x. T) Enbedding: (Ex 1, Ex 2, …, Ex. T) where Ext is the word vector of xt Bi-directional GRU: hidden states H=(h 1, h 2, …, h. M) 10

METHOD for aspect k • Multi-Aspect Self-Attention Determine attention weights αkt of each token in the review Representation of the review : sum of all hidden states 11

METHOD for aspect k • Aspect-Keywords Guided-Attention Given a list of aspect-keywords for aspect k: (gk 1, gk 2, . . . , gk. M) And enbedding as: (Egk 1, Egk 2, …, Egk. M) Each word vector is transformed into a hidden state with: Zt rt 12

METHOD for aspect k • Aspect-Keywords Guided-Attention Concatenating all hidden states into a single vector vk=[vk 1 , vk 2 , . . . , vk. M] Calculate alignment scores between encoded vectors of aspect-keywords and tokens in the review as: The guided-attention weights and vector representation of the review: 13

METHOD for aspect k • Aspect-Specific Feature Encoder We embed one hot representations of features of doctors into a continuous vector space for each aspect k as Feature of doctors : specialties, insurance plans, locations, genders and facilities 14

METHOD for aspect k • Multi-Aspect Rating Prediction Classifier to predict rating: A single layer feed-forward network with a softmax activation function Integer score from 1 to 5 15

METHOD • Loss function Score from 1 to 5 Parameter set including all weight matrices and bias vectors 16

OUTLINE l. Introduction l. Method l. Experiment l. Conclusion 17

EXPERIMENT Data. Sets ratemds-us and ratemds-ca 90% of reviews are from these two countries reviews doctors ratemds-us 1, 414, 235 385, 407 ratemds-ca 1, 252, 941 99, 719 Randomly split each dataset into training/development/testing sets at the ratio of 0. 8/0. 1. 18

EXPERIMENT Comparison with Previous Models (‘Macro’ avg F-score and MSE) 19

EXPERIMENT Comparison with Previous Models (‘Macro’ avg F-score and MSE) 20

EXPERIMENT Attention Visualization (ground-truth, predicted rating) Self-ATT weight Guided-ATT weight 21

EXPERIMENT Attention Visualization See Figure 7 Failure of attention may be caused by: u. Short Review or does not cover all aspects. u. May be mixed with unrelated sentences. u. Many keywords and phrases are ambiguous in different context. 22

OUTLINE l. Introduction l. Method l. Experiment l. Conclusion 23

CONCLUSION üWe attempted to explore content of reviews by extracting aspectkeywords with the topic modeling. üThe proposed model takes both features of doctors and aspectkeywords into consideration. üWe proposed a multi-task learning framework for the document-level multi-aspect sentiment classification. 24